A Case-Based Recommendation Approach for Market Basket Data
RECOMMENDER SYSTEMS
A Case-Based
Recommendation
Approach for Market
Basket Data
Anna Gatzioura and Miquel Sànchez-Marrè, Universitat Politècnica de
Catalunya · BarcelonaTech
R
Recently,
ecommender systems (RSs) have been used in numerous domains
to support both users in handling information overload and find-
ing adequate items to cover their needs and providers in identifying user
recommender systems needs and increasing the amount and diversity of the items they sell.1
have become an
important part of
various applications.
A novel case-based
recommendation
approach aims to
overcome the current
limitations while
providing more
insight into user
preferences and item
selection patterns.
20
The most widely used recommendation
techniques tend to recommend similar items
to those liked by a user in the past or items
that similar users liked based on the hypothesis that users have a stable behavior
over time and that similar users share similar tastes. However, these methodologies
show limited performance when new users
or less popular items appear—often called
a cold start—while others can lead to overspecialization of the recommended items.
In addition, because they tend to ignore
the underlying structure of user preferences
and don’t evaluate the hidden concepts under which items are selected, their accuracy
decreases when trading items with unequal
probability distributions.
Previous research, mainly in the fields of
market research and customer behavior analysis, has revealed the existence of patterns
that determine the structure of users’ market
baskets, which is defined as the set of items
bought by one customer in a single visit to
a store. Market basket analysis (MBA) refers to the search for meaningful associations in customer purchase data, and it aims
to discover the patterns that define the composition of those baskets.2 The market basket domain is characterized by the existence
of a large number of items and transactions
that consist of co-occurring items. More than
simply predicting whether a single item will
be liked by a user, the intention is to capture
the presence or absence of an item within
a concrete buying concept and to become
able to recommend complementary items.
Most of the current recommendation methodologies don’t take into account these cooccurrences, and the association rules (ARs)
methodology that has been used as a basis
for recommendations in such cases has several limitations—such as computational and
evaluation difficulties—when applied to large
datasets.3 There’s a need for intelligent recommendation methodologies that can generate valuable recommendations based on user
1541-1672/15/$31.00 © 2015 IEEE
Published by the IEEE Computer Society
IEEE INTELLIGENT SYSTEMS
ri1
ri2
?
Examine user-item matrix
Recommender Systems
and Recommendation
Techniques
RSs are software tools and techniques
for information retrieval and filtering
that aim to provide meaningful and
effective item recommendations to
the active user.4 The term item refers
to the type of entity (product, service,
information, and so on) being recommended; it depends on the application
area and on the specific system’s objectives. RSs usually generate a set of
(top-N) items expected to be liked by
the user or intend to predict whether a
specific item will be of interest.
The widely used recommendation
methodologies in commercial applications can be mainly divided into
collaborative filtering (CF) and content-based (CB). Figure 1 illustrates
their main differences.
In CF recommendation techniques,
items among those liked by similar
JANUARY/FEBRUARY 2015
Active user (ui)
User/item
u1
u2
uk
i1 [a11, …, a1t] i2 [a21, …, a2t]
r11
r21
in [an1, …, ant]
rn1
r12
r22
rn2
r1k
r2k
rnk
Collaborative filtering
Content-based
Item attributes (a)
Item ratings (r)
habits and purchase patterns, yet that
overcome the drawbacks of the current recommendation techniques.
Here, we present an analysis and recommendation approach for situations
in which a user’s experience depends
on the set of items selected together
more than on each item’s stand-alone
attributes. Case-based reasoning (CBR)
is particularly appropriate because the
sets of items selected together can be
adequately modeled and compared. In
addition, the performance of a casebased reasoner benefits from the existence of a large amount of data, such
as in the MBA domain. To determine
the structure of user preferences and
generate valuable recommendations,
the implemented methodology uses a
hierarchical categorization of items in
transactions. By evaluating sets of selected items, the system can identify
those items selected within concrete
concepts and recommend them in similar situations to new or existing users.
User (u)
Item (i )
Items of similar users
(through their ratings)
Similar items
(through their attributes)
Figure 1. Collaborative filtering and content-based recommendations. The main
differences are in the way that these techniques explore the user preference matrix
to extract information about user selections and generate recommendations. CF
groups users based on the similarity of their ratings while CB groups items based on
the similarity of their characteristics.
users (“neighbors”) are recommended
to the active user. A user profile is built
of the items that the user has rated
highly, thus similarities in user tastes
are deduced from previous ratings. Although widely used in commercial applications, collaborative RSs still have
to overcome scalability and cold-start
problems that limit their performance.5
On the other hand, in CB techniques, user profiles are built from
the characteristics of the items that
a user has rated highly, and the items
that he or she hasn’t yet tried are compared against them. The items with
the higher estimated possibility of being liked are then recommended.4,6
Because CB techniques rely on more
www.computer.org/intelligent
specific information about users and
items, they’re able to recommend new
items. However, they must overcome
the recommendations’ limited diversity and possible overspecialization.
Various hybrid approaches have
been proposed to leverage the
strengths of both techniques, overcome their current limitations, and
improve their recommendation accuracy.1,6 The following are the basic
recommendation techniques identified
in the literature1:
• CF (memory-based, model-based),
• CB filtering (neural networks,
probabilistic models, naïve Bayes
classifier),
21
RECOMMENDER SYSTEMS
Related Work in Case-Based Reasoning and Recommenders
C
ase-based reasoning (CBR)
is a problem-solving paraNew case description (cj)
digm closely related to the
…
aj1
human way of reasoning and acting in everyday situations when
al1
alt
facing new problems. CBR uses
RETRIEVE
New learned case (cl)
old experiences to solve new
Case base
problems, based on the following
sentence, known as the CBR asCase\attributes
RETAIN
sumption: “Similar problems have
c1
a11
a1t
K most similar cases
similar solutions.”
c2
a21
a2t
A situation experienced in the
aj1
ajt
cn
an1
ant
way that it has been captured
and learned is referred to as past/
a(j+k-1)M
a(j+k-1)1
as1
ast
previous case and is stored in
Repaired solution (cs)
Domain-user
the case base. A new situation
specific knowledge
asking for a solution forms the
description of a new/target case.
REUSE
An important part of the CBR
REVISE
methodology is its learning ability,
which comes as a natural result of
asj1
asjt
its problem-solving process: the
case base is updated each time a
Proposed solution(csj)
new experience is obtained. This
knowledge can be reused when
Figure A. Case-based reasoning (CBR) cycle. A cyclical process comprising of four
needed without implementing
processes (retrieve, reuse, revise, retain) that retrieves from the case base the most
the whole process from scratch,
similar cases to the new case and adapts their solutions to the needs of the current
or to highlight a methodology
problem to find an adequate solution.
that should be avoided in a
similar situation. Therefore, case• retain the parts of the new solution that are likely to be
based reasoners are able to improve their problem-solving
used for future purposes.
performance over time.1
The CBR solving and learning process can be described as
In addition to the knowledge obtained from previous
a cyclical process comprising four processes, known as the
cases, there’s also domain-dependent knowledge
CBR cycle (shown in Figure A), or “the four Rs”:2
supporting the CBR process.
Cases can be viewed as composed of two parts: the
• retrieve the most relevant cases,
problem description and the problem solution. Thus, cases
• reuse the knowledge provided to the new problem,
with similar descriptions are retrieved and their solutions
• revise the solution obtained, and
• knowledge-based (case-based, constraint-based),
• utility-based,
• rule-based,
• demographic, and
• other (context-aware, semantically
enhanced, hybrid).
Knowledge-based and especially
case-based recommenders have emerged
as the primary alternative to CF recommenders, intending to overcome
their shortcomings while efficiently
handling the existing information
overload. Case-based recommenders
22
implement a type of CB recommendation that relies on a structured representation of cases, usually as sets of
well-defined characteristics with their
values.7 These systems generally recommend items similar to those that
the active user has described in his or
her request (see the “Related Work in
Case-Based Reasoning and Recommenders” sidebar).
Rule-based techniques generate
item recommendations based on a set
of rules extracted from a data corpus. ARs mining refers to transaction
analysis aiming to discover interesting
www.computer.org/intelligent
hidden patterns and frequent associations among existing items, usually
expressed in the form of “if-then”
statements.3
Recently, semantic analysis, latent
factors, and probabilistic topic models arising from natural language
processing have been successfully
applied to information retrieval and
RSs, especially for tag recommendations. The basic idea is that topics are
sets of words from a given vocabulary, and documents are formed as
probability distributions over topics. These techniques show higher
IEEE INTELLIGENT SYSTEMS
are adapted to the needs of the target problem.1 The
existence of a common and structured representation of
the treated items enables case-based recommenders in
calculating the similarities and generating meaningful
recommendations of high quality—especially for new
items or to new users. Let P, Q be the subsets of problem
descriptions and solutions, respectively, then cases can be
denoted as ordered pairs c = (p, q), where p ∈ P and q ∈ Q,
while the case base in a CBR system can be defined as the
set of the known cases, C = (P, Q).
The key idea behind CBR is that similar problems
have similar solutions. Therefore, a core concept of the
CBR methodology—and a key factor of its successful
application—is the similarity measure used to identify
similar cases. These cases can have the same solution, or
their solution can be easily adapted to match the current
problem’s characteristics. Case-based recommenders follow
the general CBR cycle and rely on the CBR core concepts
of similarity and retrieval. For a case-based recommender,
the user query serves as a new problem specification, while
the available items and their descriptions form the cases in
the case base. The items to be recommended are retrieved,
based on their similarity to the user’s request. 3
Let the description of a case with t attributes be
c[a1, …, at]. To calculate the global similarity between two
cases, first the local similarities of the different attributes
have to be specified. Usually, the global similarity can be
calculated as the aggregation of the local similarities from
the following equation,4
∑ w × sim aI , aR
) = i =1 ∑i n w ( i i ) ,
i =1 i
n
(
I
Sim c , c
R
where aiI , aiR are the values of the ith attribute in the
I R
input cI and the retrieved cR cases, respectively, sim ai , ai
is the local similarity function used for their comparison,
and wi ∈ [0,1] is the corresponding weighting factor. Depending on the system’s purpose and the type of recommended items, different local similarity functions are used,
while the weighting factors may be specified by feature
(
accuracy than rule-based options
and are able to better handle sparsity
problems.8
Due to the evolution of mobile devices and the use of recommender systems in applications highly depended
on context (location, time, weather,
movement state, emotional state, and
so on), context-aware recommenders
are receiving more attention. They involve much more than grouping users or items based on their ratings or
characteristics—instead, they group
users or items associated with similar
context information.9
JANUARY/FEBRUARY 2015
)
weighting algorithms or by users as expression of their
preferences.
CBR recommenders have been mainly applied in
product recommendations, especially in electronic stores,
to support intelligent product selection by identifying the
products that best match a user’s request. 3 These systems
focus on the processes of mapping user requirements into
a proper problem specification, selecting and retrieving
items based on their similarity to the user’s request, and
their presentation, as there are few stores that enable
product customization. CBR recommenders have also
been applied in travel recommendations for both travel
services (hotels, museums, and so on) and complete
travel plans. 5 In addition, due to their case modeling
and the higher flexibility offered to decision-making
processes, CBR systems can be used in applications where
contextual and other user-specific information related to
the time of the selection must be incorporated. Another
interesting application of CBR is related to music playlists
generation. A playlist is a coherent collection of music
items of specific characteristics presented in a meaningful
sequence, thus modeling entire playlists as cases enables
identifying the relevance of songs based on their
co-occurrences.6
References
1. R. Bergmann, Introduction to Case-Based Reasoning, Univ.
Kaiserlautern, 1998.
2. R. L. de Mántaras, “Case-Based Reasoning,” Machine Learning
and Its Application, vol. 2049, 2001, pp. 127–145.
3. F. Lorenzi and F. Ricci, “Case-Based Recommender Systems: A
Unifying View,” Intelligent Techniques for Web Personalization, Springer, 2005, pp. 89–113.
4. G. Finnie and Z. Sun, “Similarity and Metrics in Case-Based
Reasoning,” Int’l J. Intelligent Systems, vol. 17, no. 3, 2002,
pp. 273–287.
5. F. Ricci et al., “ITR: A Case-Based Travel Advisory System,”
Case-Based Reasoning, Springer, vol. 2416, 2002, pp. 613–627.
6. C. Baccigalupo and E. Plaza, Case-Based Sequential Ordering
of Songs for Playlist Recommendation, LNCS 4106, T. RothBerghofer, M.H. Göker, and H.A. Güvenir, eds., Springer, 2006,
pp. 286–300.
Approach Description
The majority of recommendation techniques ignore the fact that, in many
situations, the utility a user obtains depends on the set of items selected together more than on the selected item’s
isolated attributes. Sometimes these
“relations” among items seem obvious,
while in other situations, they must be
inferred from the underlying patterns.
CF techniques evaluate only the
ratings assigned by users to items,
whereas CB techniques and the majority of case-based recommenders
focus on the characteristics of items
www.computer.org/intelligent
that a user has liked or requested. The
commonly used approach in MBA—
the ARs mining methodology—evaluates the presence or absence of items
within a transaction to extract patterns and generate recommendations
based on them. Recently, latent semantic analysis and probabilistic
topic models have been applied to RSs
for recommendations on sets of items,
evaluating the similarities of latent
topics to recommend those items.
Our proposed recommender uses
case-based reasoning (CBR) to identify and recommend the items that
23
RECOMMENDER SYSTEMS
seem more suitable for completing
a user’s buying experience provided
that he or she has already selected
some items. The system models complete transactions as cases and recommended items come from the
evaluation of those transactions. Because the cases aren’t restricted to the
user who purchased them, the developed system can generate accurate
item recommendations for joint item
selections, both for new and existing users. Having analyzed the previous transactions and identified the
concepts within which concrete items
appear, the given part of a new transaction is matched over the existing
ones to find the more adequate solution—that is, the best way to fill this
basket.
Following the formalization used
in MBA, let I = {i1, i 2 , … , il} be the
set of l distinct elements called items
that can be found in a database. Let
D be a transactional database that
contains a set of transactions T = {t 1,
t 2 , … , tm}, where each transaction tj =
{ia , ib, …} contains a subset of items
from I that were purchased together
at a specific period of time by a user
from the set of users U = {u1, u 2 , … ,
uk}. In contrast to most CBR item
recommenders7,10 that trade items
as cases and their attributes form
the case description i[a1, …, at], we
model every transaction as a case and
its items as attributes of the case c[i1,
…, ij], to capture the sets of items selected together. A case can be denoted
as an ordered pair c = (p, q), where p
= {ii1, … , iij} ∈ P is the problem description (the set of already selected
items) and q = {iil, … , ii(l+n-1)} ∈ Q is
the problem solution of size n (the set
of items that will complete this transaction), both subsets of I with p ∩ q =
∅. The case base C = (P, Q) is the set
of transactions that have been performed, where all the included items
are known.
24
The treated items aren’t (directly)
associated with quality attributes
that would either enable a contentbased comparison or affect their acceptance by users. Therefore, we view
all the items as attributes of the same
importance for the user’s buying experience. The global similarity of two
cases is the aggregated similarity of
the local similarities of the items included in them with equal weighting
factors, therefore
(
) ∑ ti =1 sim ( iiI , iiR ).
Sim c I, c R =
Currently, a lot of alternative items/
services able to cover the same, or
very similar, needs can be found,
with their differences mainly encountered in marketing characteristics
rather than in core specifications. In
addition, although the number of registered users is usually high, each one
experiences only a small percentage
of the available items. Consequently,
the resulting buying patterns are
sparse and may represent situations
that occur by chance. To identify
user habits, it’s important to identify
the specification of an item that appears within a concrete concept. For
instance, the set of items bought together before a football game, such as
soda, beer, chips, and so on, can be
seen as the “snacks for watching TV”
concept.
A preprocessing phase that analyzes items and transforms them to
a convenient representation for further comparison is important. More
than classifying items based on the
exact values of their characteristics
or the ratings assigned to them, our
approach categorizes items based on
the types of their hierarchical attributes (the categories that they belong
to). As the tree structure in Figure 2
shows, starting from the generic item
concept, we specify at each level one
more category that an item belongs
www.computer.org/intelligent
to, dividing the existing sets of items
into smaller, more coherent subsets.
As we go lower in the tree, these sets
get smaller until we reach the leaves
with the unique items. At each level,
the items that have common ancestors and aren’t differentiated at the
current level are grouped together
and treated as the “same.” For instance, the items regarded as being
the “same” item at the third level are
the items belonging to the same categories at the first two levels with
the same characteristics at the third
level. So, in a physical store at the second level, all types of milk would be
grouped together, while at the third
level, based on the type of milk, full
and semi-skimmed milk will be distinguished; at the forth level, we’ll
have distinctions based on the distribution package, and the brand name
will specify the unique item.
Because meaningful information
about the items can’t be extracted
from their initial unique IDs, the new
groups of items are labeled with appropriate “IDs/tags” to enable their
comparison. These IDs are generated
based on the categories that the items
belong to, following the path from
the tree root to its position. Having
specified the level of detail needed,
the required IDs are generated as in
Figure 2; the similarity of items in the
new and existing cases can be calculated based on those IDs. Two items
iiI , iiR from the input and a retrieved
case are thought to be similar to the
extent to which they share the same
path from the tree root to their position. Thus, their level of similarity
can be calculated from the following
equation:
(
)
sim iiI, iiR =
(
LengthOfCommonPath iiI , iiR
( )
level iiI
).
Depending on the application domain, the hierarchical attributes and
their classification may be extracted
IEEE INTELLIGENT SYSTEMS
K most similar cases
Active user / new case (cj)
ij1
ij1
?
ijt
i(j+k-1)1
Recommendations
Set of n items
i(j+k-1)t
Analyze
cases into
items
Case base
Case
Sim(cl,cj) = ∑ sim(ilk,ijk)
Similarities of
cases
Item
c1
i11
i1t
c2
i21
i2t
cn
in1
int
Items
Products
(1...N)
L1.1
L1.N
Drinks
(1...K)
sim(ilk , ijk )
L2.K
L2.1
Milk
Similarities
of items
(1...L)
L3.L
L3.1
Item
Level1
Level2
Level3
Level4
FinalID
X
L1.1
Y
L1.1
L2.1
L2.1
L3.1
L3.1
L4.M
L4.M
1.1.1.M
1.1.1.M
Z
L1.1
L2.1
L3.1
L4.1
1.1.1.1
Transform
items
Semi‐skimmed
Full-fat
(1...M)
L4.1
L4.M
Paper brick
Glass bottle
Similarity (X, Y) > Similarity (X, Z)
Brand m1 Z
W
Brand m2 X
Y Brand m3
Figure 2. Recommendation process. The similarity of cases is calculated as the aggregated similarity of the items these cases
are composed of, while the items are compared based on their hierarchical characteristics. Following the general CBR cycle, the
most similar cases are retrieved and the sets of items completing these cases form the set of recommended items.
from a proper ontology or a simpler
categorization. The described classification of items forms part of the recommendation preprocessing phase
and is used to calculate similarities between cases and to highlight the type
of items appearing in different cases.
Nevertheless, the implemented recommendation methodology doesn’t generate recommendations based on a
clustering of items.
When a new user comes, the set
of already selected items is compared to those in the existing cases,
to identify past cases with similar
descriptions and the way they were
structured. The k most similar cases
are retrieved, and the items that most
JANUARY/FEBRUARY 2015
frequently appear in their solution
parts are recommended to the user.
Therefore, even if the new case belongs to an unknown user, or if it
contains items that appear only under certain circumstances, the recommender is able to generate accurate
recommendations.
Experimental Results
and Evaluation
To evaluate the implemented recommender, we used a transactional dataset
with real market data from a European
supermarket. The number of offered
items was 102,142, with 1,057,076
transactions performed by 17,672 customers. Each transaction is associated
www.computer.org/intelligent
with the user who purchased it and the
included items. Each item, apart from
its name and unique ID, is associated
with the various categories that it belongs to (general category, item group,
and two item subgroups that, for example, would be, drink, milk, semiskimmed, glass bottle of 1 liter, brand
name). This information was used to
transform the item representation into
a proper depiction, as described above,
before generating recommendations.
Demographic and subscription information about users can also be found
in this dataset.
The recommender was tested for
different values of parameters that
affect its performance, such as the
25
RECOMMENDER SYSTEMS
Table 1. Recommendation results in terms of precision (Pr), recall (R), and f-measure (F1) for the use of level 2
and level 3 item representations.
No.
items
Association Rules
Latent Dirichlet Allocation
Collaborative Filtering
Case-Based Reasoning
Pr
R
F1
Pr
R
F1
Pr
R
F1
Pr
R
F1
5
0.226
0.089
0.128
0.418
0.084
0.139
0.242
0.205
0.222
0.446
0.177
0.253
7
0.282
0.080
0.125
0.493
0.099
0.164
0.308
0.252
0.277
0.554
0.206
0.300
9
0.337
0.075
0.123
0.531
0.106
0.177
0.366
0.296
0.327
0.629
0.226
0.333
11
0.390
0.071
0.120
0.592
0.118
0.197
0.417
0.321
0.363
0.678
0.239
0.353
5
0.137
0.040
0.062
0.086
0.017
0.029
0.072
0.064
0.068
0.318
0.110
0.163
7
0.168
0.038
0.062
0.135
0.019
0.034
0.093
0.082
0.087
0.445
0.154
0.229
Level 2
Level 3
9
0.198
0.037
0.062
0.190
0.021
0.038
0.114
0.099
0.106
0.519
0.176
0.263
11
0.225
0.036
0.062
0.197
0.018
0.033
0.140
0.119
0.129
0.570
0.185
0.279
number of similar cases retrieved (k),
the number of recommended items,
and the level of detail at which the
items are represented. To specify the
number of similar cases that would
be used, recommendations were generated using 1, 5, 10, 20, and 30 similar cases. The experimental results
show that the best option is to use
only the most similar case, thus k was
set equal to 1.
We compared the performance of
the developed recommender system
to three of the widely used techniques,
namely, AR, probabilistic topic models, and CF. However, only the first
two techniques address exactly the
same problem: recommendations of
sets of items. CF recommendations focus on the ratings users assign to items
and don’t take into account joint
item selections. But because the CF
technique is widely used in commercial applications, its results are also
presented. Item descriptions in the
transactional database don’t contain
quality attributes, so we didn’t use a
CB or a CBR item-based approach.
The Apriori algorithm was used to
extract ARs from the given transactional database, based on which the
recommendations were generated. In
addition, we used a topic model recommender (Latent Dirichlet Allocation, or LDA). In this approach, the
26
offered items are seen as words of
a vocabulary, and transactions are
treated as documents formed from a
combination of topics (item concepts)
with some probability distribution.11
In the CF approach, item selection
means the item has a high user rating.
We ran various experiments, randomly selecting each time 20 percent of the transactional database for
new cases (test set) and the rest as the
case base (training set). Using only
the most similar case (k = 1), Table 1
shows the average results for the recommendation of 5, 7, 9, and 11 items
for different representation levels. Because our intention was to evaluate
the recommender’s ability to identify
and recommend the missing items
in the transactions, information retrieval metrics (precision, recall, and
f-measure) were used for the evaluation, with our focus being on the precision value.
As you can see, the accuracy of all the
methodologies highly improves when
using more abstract descriptions. However, as the CBR recommender evaluates the degree of an item’s similarity
with the items in the target case and
not just an item’s presence or absence,
it outperforms the other recommenders at both representation levels. In
contrast, the LDA recommender evaluates similarities among item concepts
www.computer.org/intelligent
(topics), while the ARs recommender
evaluates only the presence or absence
of items within transactions to extract
the buying patterns and generate recommendations. Finally, CF recommenders
take into account only the presence of
items in the user profiles without evaluating the items’ co-occurrences within
transactions.
R
ecommender systems have become an important part of
num erous commercial applications,
enabling customers and providers in
their decision-making processes while
pursuing their buying and selling
strategies. The identification of item
selection patterns is thus of high importance. One of our approach’s main
advantages is its ability to recommend
complementary items to the already selected ones. Additionally, it can recommend less popular items and generate
recommendations to new users, reducing the cold start and the overspecialization problems of CF and CB techniques.
This work could be further extended
by incorporating a second processing
level into the recommendation methodology. At this level, additional quality characteristics of the items and
constraints related to them could be incorporated (for example, to recommend
only new items). In addition, it could be
IEEE INTELLIGENT SYSTEMS
THE AUTHORS
applied to other domains where the outcome of a user’s experience depends on
the total set of items used together. Finally we plan to examine the temporal
characteristics of the selected items.
References
1. F. Ricci, L. Rokach, and B. Shapira,
“Introduction to Recommender Systems
Handbook,” Recommender Systems Handbook, F. Ricci et al., eds.,
Springer, 2011, pp. 1–35.
2. L. Cavique, “A Scalable Algorithm for
the Market Basket Analysis,” J. Retailing and Consumer Services, vol. 14,
no. 6, 2007, pp. 400–407.
3. R. Agrawal, T. Imielin´ski, and A. Swami,
“Mining Association Rules between
Sets of Items in Large Databases,” ACM
SIGMOD Record, vol. 22, 1993,
pp. 207–216.
4. P. Melville and V. Sindhwani, “Recommender Systems,” Encyclopedia of
Machine Learning, Springer, 2010,
pp. 829-838.
5. Z. Huang, D. Zeng, and H. Chen, “A
Comparison of Collaborative-Filtering
Recommendation Algorithms for
Anna Gatzioura is a PhD student at the Universitat Politècnica de Catalunya ·
BarcelonaTech. Her research interests include intelligent decision support systems, recommender systems, machine learning, and artificial intelligence applications. Contact her
at gatzioura@cs.upc.edu.
Miquel Sànchez-Marrè is an associate professor in the computer science department at
Universitat Politècnica de Catalunya · BarcelonaTech. His research interests include casebased reasoning, machine learning, data mining, intelligent decision support systems,
integrated architectures for AI, and applications of IDSS. Sànchez-Marrè has a PhD in
computer science from the Universitat Politècnica de Catalunya · BarcelonaTech. Contact
him at miquel@cs.upc.edu.
E-commerce,” IEEE Intelligent Systems, vol. 22, no. 5, 2007, pp. 68–78.
6. G. Adomavicius and A. Tuzhilin, “Toward the Next Generation of Recommender Systems: A Survey of the Stateof-the-Art and Possible Extensions,”
IEEE Trans. Knowledge and Data Eng.,
vol. 17, no. 6, 2005, pp. 734–749.
7. F. Lorenzi and F. Ricci, “Case-Based Recommender Systems: A Unifying View,”
Intelligent Techniques for Web Personalization, Springer, 2005, pp. 89–113.
8. M. Steyvers and T. Griffiths, “Probabilistic Topic Models,” Handbook of
Latent Semantic Analysis, vol. 427, 2007,
pp. 424–440.
9. J.H. Su et al., “Music Recommendation
Using Content and Context Information
Mining,” IEEE Intelligent Systems,
vol. 25, no.1, 2010, pp. 16–26.
10. D. Bridge et al., “Case-Based Recommender Systems,” Knowledge Eng.
Rev., vol. 20, no. 3, 2006, pp. 315–320.
11. K. Christidis, D. Apostolou, and G.
Mentzas, “Exploring Customer Preferences with Probabilistic Topic Models,”
Proc. European Conf. Machine Learning, 2010; www.ke.tu-darmstadt.de/
events/PL-10/papers/3-Christidis.pdf.
Selected CS articles and columns
are also available for free at
http://ComputingNow.computer.org.
Engineering and Applying the Internet
IEEE Internet Computing reports emerging tools,
technologies, and applications implemented through the
Internet to support a worldwide computing environment.
For submission information and author guidelines,
please visit www.computer.org/internet/author.htm
JANUARY/FEBRUARY 2015
www.computer.org/intelligent
27
A Case-Based
Recommendation
Approach for Market
Basket Data
Anna Gatzioura and Miquel Sànchez-Marrè, Universitat Politècnica de
Catalunya · BarcelonaTech
R
Recently,
ecommender systems (RSs) have been used in numerous domains
to support both users in handling information overload and find-
ing adequate items to cover their needs and providers in identifying user
recommender systems needs and increasing the amount and diversity of the items they sell.1
have become an
important part of
various applications.
A novel case-based
recommendation
approach aims to
overcome the current
limitations while
providing more
insight into user
preferences and item
selection patterns.
20
The most widely used recommendation
techniques tend to recommend similar items
to those liked by a user in the past or items
that similar users liked based on the hypothesis that users have a stable behavior
over time and that similar users share similar tastes. However, these methodologies
show limited performance when new users
or less popular items appear—often called
a cold start—while others can lead to overspecialization of the recommended items.
In addition, because they tend to ignore
the underlying structure of user preferences
and don’t evaluate the hidden concepts under which items are selected, their accuracy
decreases when trading items with unequal
probability distributions.
Previous research, mainly in the fields of
market research and customer behavior analysis, has revealed the existence of patterns
that determine the structure of users’ market
baskets, which is defined as the set of items
bought by one customer in a single visit to
a store. Market basket analysis (MBA) refers to the search for meaningful associations in customer purchase data, and it aims
to discover the patterns that define the composition of those baskets.2 The market basket domain is characterized by the existence
of a large number of items and transactions
that consist of co-occurring items. More than
simply predicting whether a single item will
be liked by a user, the intention is to capture
the presence or absence of an item within
a concrete buying concept and to become
able to recommend complementary items.
Most of the current recommendation methodologies don’t take into account these cooccurrences, and the association rules (ARs)
methodology that has been used as a basis
for recommendations in such cases has several limitations—such as computational and
evaluation difficulties—when applied to large
datasets.3 There’s a need for intelligent recommendation methodologies that can generate valuable recommendations based on user
1541-1672/15/$31.00 © 2015 IEEE
Published by the IEEE Computer Society
IEEE INTELLIGENT SYSTEMS
ri1
ri2
?
Examine user-item matrix
Recommender Systems
and Recommendation
Techniques
RSs are software tools and techniques
for information retrieval and filtering
that aim to provide meaningful and
effective item recommendations to
the active user.4 The term item refers
to the type of entity (product, service,
information, and so on) being recommended; it depends on the application
area and on the specific system’s objectives. RSs usually generate a set of
(top-N) items expected to be liked by
the user or intend to predict whether a
specific item will be of interest.
The widely used recommendation
methodologies in commercial applications can be mainly divided into
collaborative filtering (CF) and content-based (CB). Figure 1 illustrates
their main differences.
In CF recommendation techniques,
items among those liked by similar
JANUARY/FEBRUARY 2015
Active user (ui)
User/item
u1
u2
uk
i1 [a11, …, a1t] i2 [a21, …, a2t]
r11
r21
in [an1, …, ant]
rn1
r12
r22
rn2
r1k
r2k
rnk
Collaborative filtering
Content-based
Item attributes (a)
Item ratings (r)
habits and purchase patterns, yet that
overcome the drawbacks of the current recommendation techniques.
Here, we present an analysis and recommendation approach for situations
in which a user’s experience depends
on the set of items selected together
more than on each item’s stand-alone
attributes. Case-based reasoning (CBR)
is particularly appropriate because the
sets of items selected together can be
adequately modeled and compared. In
addition, the performance of a casebased reasoner benefits from the existence of a large amount of data, such
as in the MBA domain. To determine
the structure of user preferences and
generate valuable recommendations,
the implemented methodology uses a
hierarchical categorization of items in
transactions. By evaluating sets of selected items, the system can identify
those items selected within concrete
concepts and recommend them in similar situations to new or existing users.
User (u)
Item (i )
Items of similar users
(through their ratings)
Similar items
(through their attributes)
Figure 1. Collaborative filtering and content-based recommendations. The main
differences are in the way that these techniques explore the user preference matrix
to extract information about user selections and generate recommendations. CF
groups users based on the similarity of their ratings while CB groups items based on
the similarity of their characteristics.
users (“neighbors”) are recommended
to the active user. A user profile is built
of the items that the user has rated
highly, thus similarities in user tastes
are deduced from previous ratings. Although widely used in commercial applications, collaborative RSs still have
to overcome scalability and cold-start
problems that limit their performance.5
On the other hand, in CB techniques, user profiles are built from
the characteristics of the items that
a user has rated highly, and the items
that he or she hasn’t yet tried are compared against them. The items with
the higher estimated possibility of being liked are then recommended.4,6
Because CB techniques rely on more
www.computer.org/intelligent
specific information about users and
items, they’re able to recommend new
items. However, they must overcome
the recommendations’ limited diversity and possible overspecialization.
Various hybrid approaches have
been proposed to leverage the
strengths of both techniques, overcome their current limitations, and
improve their recommendation accuracy.1,6 The following are the basic
recommendation techniques identified
in the literature1:
• CF (memory-based, model-based),
• CB filtering (neural networks,
probabilistic models, naïve Bayes
classifier),
21
RECOMMENDER SYSTEMS
Related Work in Case-Based Reasoning and Recommenders
C
ase-based reasoning (CBR)
is a problem-solving paraNew case description (cj)
digm closely related to the
…
aj1
human way of reasoning and acting in everyday situations when
al1
alt
facing new problems. CBR uses
RETRIEVE
New learned case (cl)
old experiences to solve new
Case base
problems, based on the following
sentence, known as the CBR asCase\attributes
RETAIN
sumption: “Similar problems have
c1
a11
a1t
K most similar cases
similar solutions.”
c2
a21
a2t
A situation experienced in the
aj1
ajt
cn
an1
ant
way that it has been captured
and learned is referred to as past/
a(j+k-1)M
a(j+k-1)1
as1
ast
previous case and is stored in
Repaired solution (cs)
Domain-user
the case base. A new situation
specific knowledge
asking for a solution forms the
description of a new/target case.
REUSE
An important part of the CBR
REVISE
methodology is its learning ability,
which comes as a natural result of
asj1
asjt
its problem-solving process: the
case base is updated each time a
Proposed solution(csj)
new experience is obtained. This
knowledge can be reused when
Figure A. Case-based reasoning (CBR) cycle. A cyclical process comprising of four
needed without implementing
processes (retrieve, reuse, revise, retain) that retrieves from the case base the most
the whole process from scratch,
similar cases to the new case and adapts their solutions to the needs of the current
or to highlight a methodology
problem to find an adequate solution.
that should be avoided in a
similar situation. Therefore, case• retain the parts of the new solution that are likely to be
based reasoners are able to improve their problem-solving
used for future purposes.
performance over time.1
The CBR solving and learning process can be described as
In addition to the knowledge obtained from previous
a cyclical process comprising four processes, known as the
cases, there’s also domain-dependent knowledge
CBR cycle (shown in Figure A), or “the four Rs”:2
supporting the CBR process.
Cases can be viewed as composed of two parts: the
• retrieve the most relevant cases,
problem description and the problem solution. Thus, cases
• reuse the knowledge provided to the new problem,
with similar descriptions are retrieved and their solutions
• revise the solution obtained, and
• knowledge-based (case-based, constraint-based),
• utility-based,
• rule-based,
• demographic, and
• other (context-aware, semantically
enhanced, hybrid).
Knowledge-based and especially
case-based recommenders have emerged
as the primary alternative to CF recommenders, intending to overcome
their shortcomings while efficiently
handling the existing information
overload. Case-based recommenders
22
implement a type of CB recommendation that relies on a structured representation of cases, usually as sets of
well-defined characteristics with their
values.7 These systems generally recommend items similar to those that
the active user has described in his or
her request (see the “Related Work in
Case-Based Reasoning and Recommenders” sidebar).
Rule-based techniques generate
item recommendations based on a set
of rules extracted from a data corpus. ARs mining refers to transaction
analysis aiming to discover interesting
www.computer.org/intelligent
hidden patterns and frequent associations among existing items, usually
expressed in the form of “if-then”
statements.3
Recently, semantic analysis, latent
factors, and probabilistic topic models arising from natural language
processing have been successfully
applied to information retrieval and
RSs, especially for tag recommendations. The basic idea is that topics are
sets of words from a given vocabulary, and documents are formed as
probability distributions over topics. These techniques show higher
IEEE INTELLIGENT SYSTEMS
are adapted to the needs of the target problem.1 The
existence of a common and structured representation of
the treated items enables case-based recommenders in
calculating the similarities and generating meaningful
recommendations of high quality—especially for new
items or to new users. Let P, Q be the subsets of problem
descriptions and solutions, respectively, then cases can be
denoted as ordered pairs c = (p, q), where p ∈ P and q ∈ Q,
while the case base in a CBR system can be defined as the
set of the known cases, C = (P, Q).
The key idea behind CBR is that similar problems
have similar solutions. Therefore, a core concept of the
CBR methodology—and a key factor of its successful
application—is the similarity measure used to identify
similar cases. These cases can have the same solution, or
their solution can be easily adapted to match the current
problem’s characteristics. Case-based recommenders follow
the general CBR cycle and rely on the CBR core concepts
of similarity and retrieval. For a case-based recommender,
the user query serves as a new problem specification, while
the available items and their descriptions form the cases in
the case base. The items to be recommended are retrieved,
based on their similarity to the user’s request. 3
Let the description of a case with t attributes be
c[a1, …, at]. To calculate the global similarity between two
cases, first the local similarities of the different attributes
have to be specified. Usually, the global similarity can be
calculated as the aggregation of the local similarities from
the following equation,4
∑ w × sim aI , aR
) = i =1 ∑i n w ( i i ) ,
i =1 i
n
(
I
Sim c , c
R
where aiI , aiR are the values of the ith attribute in the
I R
input cI and the retrieved cR cases, respectively, sim ai , ai
is the local similarity function used for their comparison,
and wi ∈ [0,1] is the corresponding weighting factor. Depending on the system’s purpose and the type of recommended items, different local similarity functions are used,
while the weighting factors may be specified by feature
(
accuracy than rule-based options
and are able to better handle sparsity
problems.8
Due to the evolution of mobile devices and the use of recommender systems in applications highly depended
on context (location, time, weather,
movement state, emotional state, and
so on), context-aware recommenders
are receiving more attention. They involve much more than grouping users or items based on their ratings or
characteristics—instead, they group
users or items associated with similar
context information.9
JANUARY/FEBRUARY 2015
)
weighting algorithms or by users as expression of their
preferences.
CBR recommenders have been mainly applied in
product recommendations, especially in electronic stores,
to support intelligent product selection by identifying the
products that best match a user’s request. 3 These systems
focus on the processes of mapping user requirements into
a proper problem specification, selecting and retrieving
items based on their similarity to the user’s request, and
their presentation, as there are few stores that enable
product customization. CBR recommenders have also
been applied in travel recommendations for both travel
services (hotels, museums, and so on) and complete
travel plans. 5 In addition, due to their case modeling
and the higher flexibility offered to decision-making
processes, CBR systems can be used in applications where
contextual and other user-specific information related to
the time of the selection must be incorporated. Another
interesting application of CBR is related to music playlists
generation. A playlist is a coherent collection of music
items of specific characteristics presented in a meaningful
sequence, thus modeling entire playlists as cases enables
identifying the relevance of songs based on their
co-occurrences.6
References
1. R. Bergmann, Introduction to Case-Based Reasoning, Univ.
Kaiserlautern, 1998.
2. R. L. de Mántaras, “Case-Based Reasoning,” Machine Learning
and Its Application, vol. 2049, 2001, pp. 127–145.
3. F. Lorenzi and F. Ricci, “Case-Based Recommender Systems: A
Unifying View,” Intelligent Techniques for Web Personalization, Springer, 2005, pp. 89–113.
4. G. Finnie and Z. Sun, “Similarity and Metrics in Case-Based
Reasoning,” Int’l J. Intelligent Systems, vol. 17, no. 3, 2002,
pp. 273–287.
5. F. Ricci et al., “ITR: A Case-Based Travel Advisory System,”
Case-Based Reasoning, Springer, vol. 2416, 2002, pp. 613–627.
6. C. Baccigalupo and E. Plaza, Case-Based Sequential Ordering
of Songs for Playlist Recommendation, LNCS 4106, T. RothBerghofer, M.H. Göker, and H.A. Güvenir, eds., Springer, 2006,
pp. 286–300.
Approach Description
The majority of recommendation techniques ignore the fact that, in many
situations, the utility a user obtains depends on the set of items selected together more than on the selected item’s
isolated attributes. Sometimes these
“relations” among items seem obvious,
while in other situations, they must be
inferred from the underlying patterns.
CF techniques evaluate only the
ratings assigned by users to items,
whereas CB techniques and the majority of case-based recommenders
focus on the characteristics of items
www.computer.org/intelligent
that a user has liked or requested. The
commonly used approach in MBA—
the ARs mining methodology—evaluates the presence or absence of items
within a transaction to extract patterns and generate recommendations
based on them. Recently, latent semantic analysis and probabilistic
topic models have been applied to RSs
for recommendations on sets of items,
evaluating the similarities of latent
topics to recommend those items.
Our proposed recommender uses
case-based reasoning (CBR) to identify and recommend the items that
23
RECOMMENDER SYSTEMS
seem more suitable for completing
a user’s buying experience provided
that he or she has already selected
some items. The system models complete transactions as cases and recommended items come from the
evaluation of those transactions. Because the cases aren’t restricted to the
user who purchased them, the developed system can generate accurate
item recommendations for joint item
selections, both for new and existing users. Having analyzed the previous transactions and identified the
concepts within which concrete items
appear, the given part of a new transaction is matched over the existing
ones to find the more adequate solution—that is, the best way to fill this
basket.
Following the formalization used
in MBA, let I = {i1, i 2 , … , il} be the
set of l distinct elements called items
that can be found in a database. Let
D be a transactional database that
contains a set of transactions T = {t 1,
t 2 , … , tm}, where each transaction tj =
{ia , ib, …} contains a subset of items
from I that were purchased together
at a specific period of time by a user
from the set of users U = {u1, u 2 , … ,
uk}. In contrast to most CBR item
recommenders7,10 that trade items
as cases and their attributes form
the case description i[a1, …, at], we
model every transaction as a case and
its items as attributes of the case c[i1,
…, ij], to capture the sets of items selected together. A case can be denoted
as an ordered pair c = (p, q), where p
= {ii1, … , iij} ∈ P is the problem description (the set of already selected
items) and q = {iil, … , ii(l+n-1)} ∈ Q is
the problem solution of size n (the set
of items that will complete this transaction), both subsets of I with p ∩ q =
∅. The case base C = (P, Q) is the set
of transactions that have been performed, where all the included items
are known.
24
The treated items aren’t (directly)
associated with quality attributes
that would either enable a contentbased comparison or affect their acceptance by users. Therefore, we view
all the items as attributes of the same
importance for the user’s buying experience. The global similarity of two
cases is the aggregated similarity of
the local similarities of the items included in them with equal weighting
factors, therefore
(
) ∑ ti =1 sim ( iiI , iiR ).
Sim c I, c R =
Currently, a lot of alternative items/
services able to cover the same, or
very similar, needs can be found,
with their differences mainly encountered in marketing characteristics
rather than in core specifications. In
addition, although the number of registered users is usually high, each one
experiences only a small percentage
of the available items. Consequently,
the resulting buying patterns are
sparse and may represent situations
that occur by chance. To identify
user habits, it’s important to identify
the specification of an item that appears within a concrete concept. For
instance, the set of items bought together before a football game, such as
soda, beer, chips, and so on, can be
seen as the “snacks for watching TV”
concept.
A preprocessing phase that analyzes items and transforms them to
a convenient representation for further comparison is important. More
than classifying items based on the
exact values of their characteristics
or the ratings assigned to them, our
approach categorizes items based on
the types of their hierarchical attributes (the categories that they belong
to). As the tree structure in Figure 2
shows, starting from the generic item
concept, we specify at each level one
more category that an item belongs
www.computer.org/intelligent
to, dividing the existing sets of items
into smaller, more coherent subsets.
As we go lower in the tree, these sets
get smaller until we reach the leaves
with the unique items. At each level,
the items that have common ancestors and aren’t differentiated at the
current level are grouped together
and treated as the “same.” For instance, the items regarded as being
the “same” item at the third level are
the items belonging to the same categories at the first two levels with
the same characteristics at the third
level. So, in a physical store at the second level, all types of milk would be
grouped together, while at the third
level, based on the type of milk, full
and semi-skimmed milk will be distinguished; at the forth level, we’ll
have distinctions based on the distribution package, and the brand name
will specify the unique item.
Because meaningful information
about the items can’t be extracted
from their initial unique IDs, the new
groups of items are labeled with appropriate “IDs/tags” to enable their
comparison. These IDs are generated
based on the categories that the items
belong to, following the path from
the tree root to its position. Having
specified the level of detail needed,
the required IDs are generated as in
Figure 2; the similarity of items in the
new and existing cases can be calculated based on those IDs. Two items
iiI , iiR from the input and a retrieved
case are thought to be similar to the
extent to which they share the same
path from the tree root to their position. Thus, their level of similarity
can be calculated from the following
equation:
(
)
sim iiI, iiR =
(
LengthOfCommonPath iiI , iiR
( )
level iiI
).
Depending on the application domain, the hierarchical attributes and
their classification may be extracted
IEEE INTELLIGENT SYSTEMS
K most similar cases
Active user / new case (cj)
ij1
ij1
?
ijt
i(j+k-1)1
Recommendations
Set of n items
i(j+k-1)t
Analyze
cases into
items
Case base
Case
Sim(cl,cj) = ∑ sim(ilk,ijk)
Similarities of
cases
Item
c1
i11
i1t
c2
i21
i2t
cn
in1
int
Items
Products
(1...N)
L1.1
L1.N
Drinks
(1...K)
sim(ilk , ijk )
L2.K
L2.1
Milk
Similarities
of items
(1...L)
L3.L
L3.1
Item
Level1
Level2
Level3
Level4
FinalID
X
L1.1
Y
L1.1
L2.1
L2.1
L3.1
L3.1
L4.M
L4.M
1.1.1.M
1.1.1.M
Z
L1.1
L2.1
L3.1
L4.1
1.1.1.1
Transform
items
Semi‐skimmed
Full-fat
(1...M)
L4.1
L4.M
Paper brick
Glass bottle
Similarity (X, Y) > Similarity (X, Z)
Brand m1 Z
W
Brand m2 X
Y Brand m3
Figure 2. Recommendation process. The similarity of cases is calculated as the aggregated similarity of the items these cases
are composed of, while the items are compared based on their hierarchical characteristics. Following the general CBR cycle, the
most similar cases are retrieved and the sets of items completing these cases form the set of recommended items.
from a proper ontology or a simpler
categorization. The described classification of items forms part of the recommendation preprocessing phase
and is used to calculate similarities between cases and to highlight the type
of items appearing in different cases.
Nevertheless, the implemented recommendation methodology doesn’t generate recommendations based on a
clustering of items.
When a new user comes, the set
of already selected items is compared to those in the existing cases,
to identify past cases with similar
descriptions and the way they were
structured. The k most similar cases
are retrieved, and the items that most
JANUARY/FEBRUARY 2015
frequently appear in their solution
parts are recommended to the user.
Therefore, even if the new case belongs to an unknown user, or if it
contains items that appear only under certain circumstances, the recommender is able to generate accurate
recommendations.
Experimental Results
and Evaluation
To evaluate the implemented recommender, we used a transactional dataset
with real market data from a European
supermarket. The number of offered
items was 102,142, with 1,057,076
transactions performed by 17,672 customers. Each transaction is associated
www.computer.org/intelligent
with the user who purchased it and the
included items. Each item, apart from
its name and unique ID, is associated
with the various categories that it belongs to (general category, item group,
and two item subgroups that, for example, would be, drink, milk, semiskimmed, glass bottle of 1 liter, brand
name). This information was used to
transform the item representation into
a proper depiction, as described above,
before generating recommendations.
Demographic and subscription information about users can also be found
in this dataset.
The recommender was tested for
different values of parameters that
affect its performance, such as the
25
RECOMMENDER SYSTEMS
Table 1. Recommendation results in terms of precision (Pr), recall (R), and f-measure (F1) for the use of level 2
and level 3 item representations.
No.
items
Association Rules
Latent Dirichlet Allocation
Collaborative Filtering
Case-Based Reasoning
Pr
R
F1
Pr
R
F1
Pr
R
F1
Pr
R
F1
5
0.226
0.089
0.128
0.418
0.084
0.139
0.242
0.205
0.222
0.446
0.177
0.253
7
0.282
0.080
0.125
0.493
0.099
0.164
0.308
0.252
0.277
0.554
0.206
0.300
9
0.337
0.075
0.123
0.531
0.106
0.177
0.366
0.296
0.327
0.629
0.226
0.333
11
0.390
0.071
0.120
0.592
0.118
0.197
0.417
0.321
0.363
0.678
0.239
0.353
5
0.137
0.040
0.062
0.086
0.017
0.029
0.072
0.064
0.068
0.318
0.110
0.163
7
0.168
0.038
0.062
0.135
0.019
0.034
0.093
0.082
0.087
0.445
0.154
0.229
Level 2
Level 3
9
0.198
0.037
0.062
0.190
0.021
0.038
0.114
0.099
0.106
0.519
0.176
0.263
11
0.225
0.036
0.062
0.197
0.018
0.033
0.140
0.119
0.129
0.570
0.185
0.279
number of similar cases retrieved (k),
the number of recommended items,
and the level of detail at which the
items are represented. To specify the
number of similar cases that would
be used, recommendations were generated using 1, 5, 10, 20, and 30 similar cases. The experimental results
show that the best option is to use
only the most similar case, thus k was
set equal to 1.
We compared the performance of
the developed recommender system
to three of the widely used techniques,
namely, AR, probabilistic topic models, and CF. However, only the first
two techniques address exactly the
same problem: recommendations of
sets of items. CF recommendations focus on the ratings users assign to items
and don’t take into account joint
item selections. But because the CF
technique is widely used in commercial applications, its results are also
presented. Item descriptions in the
transactional database don’t contain
quality attributes, so we didn’t use a
CB or a CBR item-based approach.
The Apriori algorithm was used to
extract ARs from the given transactional database, based on which the
recommendations were generated. In
addition, we used a topic model recommender (Latent Dirichlet Allocation, or LDA). In this approach, the
26
offered items are seen as words of
a vocabulary, and transactions are
treated as documents formed from a
combination of topics (item concepts)
with some probability distribution.11
In the CF approach, item selection
means the item has a high user rating.
We ran various experiments, randomly selecting each time 20 percent of the transactional database for
new cases (test set) and the rest as the
case base (training set). Using only
the most similar case (k = 1), Table 1
shows the average results for the recommendation of 5, 7, 9, and 11 items
for different representation levels. Because our intention was to evaluate
the recommender’s ability to identify
and recommend the missing items
in the transactions, information retrieval metrics (precision, recall, and
f-measure) were used for the evaluation, with our focus being on the precision value.
As you can see, the accuracy of all the
methodologies highly improves when
using more abstract descriptions. However, as the CBR recommender evaluates the degree of an item’s similarity
with the items in the target case and
not just an item’s presence or absence,
it outperforms the other recommenders at both representation levels. In
contrast, the LDA recommender evaluates similarities among item concepts
www.computer.org/intelligent
(topics), while the ARs recommender
evaluates only the presence or absence
of items within transactions to extract
the buying patterns and generate recommendations. Finally, CF recommenders
take into account only the presence of
items in the user profiles without evaluating the items’ co-occurrences within
transactions.
R
ecommender systems have become an important part of
num erous commercial applications,
enabling customers and providers in
their decision-making processes while
pursuing their buying and selling
strategies. The identification of item
selection patterns is thus of high importance. One of our approach’s main
advantages is its ability to recommend
complementary items to the already selected ones. Additionally, it can recommend less popular items and generate
recommendations to new users, reducing the cold start and the overspecialization problems of CF and CB techniques.
This work could be further extended
by incorporating a second processing
level into the recommendation methodology. At this level, additional quality characteristics of the items and
constraints related to them could be incorporated (for example, to recommend
only new items). In addition, it could be
IEEE INTELLIGENT SYSTEMS
THE AUTHORS
applied to other domains where the outcome of a user’s experience depends on
the total set of items used together. Finally we plan to examine the temporal
characteristics of the selected items.
References
1. F. Ricci, L. Rokach, and B. Shapira,
“Introduction to Recommender Systems
Handbook,” Recommender Systems Handbook, F. Ricci et al., eds.,
Springer, 2011, pp. 1–35.
2. L. Cavique, “A Scalable Algorithm for
the Market Basket Analysis,” J. Retailing and Consumer Services, vol. 14,
no. 6, 2007, pp. 400–407.
3. R. Agrawal, T. Imielin´ski, and A. Swami,
“Mining Association Rules between
Sets of Items in Large Databases,” ACM
SIGMOD Record, vol. 22, 1993,
pp. 207–216.
4. P. Melville and V. Sindhwani, “Recommender Systems,” Encyclopedia of
Machine Learning, Springer, 2010,
pp. 829-838.
5. Z. Huang, D. Zeng, and H. Chen, “A
Comparison of Collaborative-Filtering
Recommendation Algorithms for
Anna Gatzioura is a PhD student at the Universitat Politècnica de Catalunya ·
BarcelonaTech. Her research interests include intelligent decision support systems, recommender systems, machine learning, and artificial intelligence applications. Contact her
at gatzioura@cs.upc.edu.
Miquel Sànchez-Marrè is an associate professor in the computer science department at
Universitat Politècnica de Catalunya · BarcelonaTech. His research interests include casebased reasoning, machine learning, data mining, intelligent decision support systems,
integrated architectures for AI, and applications of IDSS. Sànchez-Marrè has a PhD in
computer science from the Universitat Politècnica de Catalunya · BarcelonaTech. Contact
him at miquel@cs.upc.edu.
E-commerce,” IEEE Intelligent Systems, vol. 22, no. 5, 2007, pp. 68–78.
6. G. Adomavicius and A. Tuzhilin, “Toward the Next Generation of Recommender Systems: A Survey of the Stateof-the-Art and Possible Extensions,”
IEEE Trans. Knowledge and Data Eng.,
vol. 17, no. 6, 2005, pp. 734–749.
7. F. Lorenzi and F. Ricci, “Case-Based Recommender Systems: A Unifying View,”
Intelligent Techniques for Web Personalization, Springer, 2005, pp. 89–113.
8. M. Steyvers and T. Griffiths, “Probabilistic Topic Models,” Handbook of
Latent Semantic Analysis, vol. 427, 2007,
pp. 424–440.
9. J.H. Su et al., “Music Recommendation
Using Content and Context Information
Mining,” IEEE Intelligent Systems,
vol. 25, no.1, 2010, pp. 16–26.
10. D. Bridge et al., “Case-Based Recommender Systems,” Knowledge Eng.
Rev., vol. 20, no. 3, 2006, pp. 315–320.
11. K. Christidis, D. Apostolou, and G.
Mentzas, “Exploring Customer Preferences with Probabilistic Topic Models,”
Proc. European Conf. Machine Learning, 2010; www.ke.tu-darmstadt.de/
events/PL-10/papers/3-Christidis.pdf.
Selected CS articles and columns
are also available for free at
http://ComputingNow.computer.org.
Engineering and Applying the Internet
IEEE Internet Computing reports emerging tools,
technologies, and applications implemented through the
Internet to support a worldwide computing environment.
For submission information and author guidelines,
please visit www.computer.org/internet/author.htm
JANUARY/FEBRUARY 2015
www.computer.org/intelligent
27